Researchers at Chalmers University of Technology in Gothenburg, Sweden, have developed an EV charging algorithm that, in simulation, extended the useful life of a lithium-ion battery cell by nearly 23 percent compared with the method used in virtually every electric vehicle on the road today. The catch: the work has been validated only on a single simulated cell, never on a real battery, and never on the kind of multi-cell pack that actually powers a car. Whether it will work in the physical world remains an open question. This is early-stage research, not a system approaching your driveway.
That said, the data from those simulations really are tantalizing. The charging-time penalty the new method imposes could be less than three seconds per session—a pretty tolerable penalty for a potential gain of two or three extra years of useful battery life. One more thing to know upfront: the approach applies only to Level 1 and Level 2 AC charging—the kind done at home or at a workplace charger. It does not apply to Level 3 DC fast charging at public stations, which operates under different electrical constraints. But for drivers who do most of their charging overnight at home, which covers the vast majority of EV owners, the potential benefit could be significant.
The paper, “Lifelong Reinforcement Learning for Health-Aware Fast Charging of Lithium-ion Batteries,” was published last month in * IEEE Transactions on *. It describes a controller that adjusts its charging behavior from session to session as a battery ages, rather than applying the same fixed voltage ceiling to a degraded five-year-old cell that it would to a new one. The standard industry protocol —
[Transportation](https://spectrum.ieee.org/tag/transportation)
[Electrification](https://spectrum.ieee.org/tag/electrification)
constant-current, constant-voltagecharging — was designed around new cells and never updated as engineers came to better understand how aging changes what a cell can safely tolerate during charging. Older cells have higher internal resistance and a narrower safe-operating window. Charging them as though they’re brand new accelerates their decline.
The Chalmers controller uses the battery’s real-time state of health — essentially a measure of how much capacity the cell retains compared to when it was new — as a live input. As health declines, the algorithm eases up, trading a few extra seconds per charging session for substantially less stress on the cell.
Trading Off Battery Protection and Charge Time #
To find a good tradeoff between protecting the battery and keeping charge times as short as possible, the researchers used a machine-learning technique that discovers good strategies through trial and error across thousands of simulated charging cycles. The resulting controller is deliberately simple: it takes two inputs—voltage and state of health — and produces one output: a current level during charging. All learning happens before the software is deployed. Once installed, the controller simply churns through calculations established during the training. . The researchers say their algorithms could reach existing vehicles as an over-the-air software update to the vehicles’ existing battery management systems, with no new hardware required.
A battery’s state of health declines at a different rate depending on the current and voltage profiles used to recharge it. Shown here are four different charging profiles, and the expected number of charging cycles that the battery could withstand before reaching a specified state of health. CCCV and CCCV-V are typical profiles used today.Changfu Zou and Meng Yuan
That deployment path is what gives the work potential relevance beyond the lab. Battery degradation is the one of the main expenses for the fleet operators who depend on EVs: taxi companies, delivery services, transit buses. For those vehicles, which are charged multiple times daily, a 23-percent life extension would meaningfully change their economics.
Getting there requires several steps that haven’t been taken. The algorithm will need to be tested against real cells, a process that takes months of physical cycling. It will then need to be validated at pack level — a production EV contains hundreds or thousands of cells, each aging at a slightly different rate. Managing a pack is a very different problem from optimizing a single cell.
The simulation also assumed a constant operating temperature of 25 degrees Celsius. Meng Yuan, the paper’s lead author and now an assistant professor at Victoria University of Wellington, says the research team chose a specific temperature because of battery management systems’ ability to hold temperature within a narrow range. But in the real world, he acknowledges, temperature changes are an important variable. . He says the team’s follow-on work will also include adapting the controller to battery chemistries other than the one tested. That, he notes, will require additional characterization work for each.
Yuan says that industry contacts have reached out since the paper’s publication, though the team has not yet engaged formally with automakers or regulators. The simulation has established a clear proof of concept. The hardware work that would confirm or complicate it has not yet begun.
*“Lifelong Reinforcement Learning for Health-Aware Fast Charging of Lithium-ion Batteries” is published in IEEE Transactions on Transportation Electrification.
Willie D. Jones Willie Jones is an associate editor at IEEE Spectrum. In addition to editing and planning daily coverage, he manages several of Spectrum's newsletters and contributes regularly to the monthly Big Picture section that appears in the print edition.